Self Organizing Neural Networks perform different from statistical k-means clustering
Alfred Ultsch · 2003
If Self Organizing Feature Maps (SOFMs) are enhanced by specially designed visualization algorithms like U-Matrix methods, they can be used for clustering. A common belief is, that the clustering abilities of this type of Artificial Neural Networks is identical or at least similar to the k-means algorithm used in statistics. In this paper we show by means of an nonlinear separable dataset, that the clustering abilities of SOFMs are quite different from the k-means algorithm. SOFMs were able to recognize clusters in a dataset where other statistical algorithms failed to produce meaningful clusters. The dataset used, called chainlink, may serve as a “benchmark” to compare clustering abilities of different algorithms.